Scholarlink

<p>Recent years have seen a significant rise in the incidence of skin cancer around the globe. However, there are not a lot of training datasets that are available, and you need a good model that can detect severe skin cancers like melanomas at an early stage. As a result, automating the categorization of skin cancer is not a simple task. Recent efforts to apply deep learning techniques for binary classification (determining if a lesion is cancerous or benign) or multiclass classification (identifying several categories of skin illnesses, such as cancer, benign lesions, and others) have shown positive results. Both of these classification methods are used to identify different types of skin disorders. In spite of the fact that the current detection rate is somewhere around 90%, there is still potential for improvement because relatively few studies have been carried out with the specific intention of detecting melanoma, which is a form of skin cancer. In addition, there has been no previous research that has examined the robustness of the models by using different datasets. This is because all of the studies have utilized the same dataset for both training and testing purposes. Furthermore, the Kaggle Skin Disease Dataset is utilized in this investigation for binary and multiclass melanoma classification. ResNet-50 and MobileNetV2 have been employed, and then HAM10000 is used to test the superior classifiers’ reliability. ResNet-50 and MobileNetV2 are trained with the Skin Disease Dataset and HAM10000 to detect melanoma better. Next, the classifiers are tested using ISIC2018, an independent test dataset. The results showed that ResNet-50 had the highest classification accuracy at 99.8%. This demonstrates that the binary classifier is superior to the multiclass classifier. Both binary models were only 50% correct when they were tested for robustness. On the other hand, ResNet-50 and MobileNetV2, both of which were trained on the pooled dataset, achieved a classification success rate of 98.8%. In contrast to the classification accuracy of 50% that was achieved through training with a single dataset, the robustness of ResNet-50 was tested, and the results showed that it achieved an accuracy of 85%.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing melanoma detection through multiple datasets integration and robust deep learning

  • 期刊论文

  • 发表日期：

    2025年12月23日

查看全文
  • Ali Al-Yousef,
  • Moy’awiah A. Al-Shannaq,
  • Ahmad A. Saifan,
  • Rami Mohawesh

摘要

Recent years have seen a significant rise in the incidence of skin cancer around the globe. However, there are not a lot of training datasets that are available, and you need a good model that can detect severe skin cancers like melanomas at an early stage. As a result, automating the categorization of skin cancer is not a simple task. Recent efforts to apply deep learning techniques for binary classification (determining if a lesion is cancerous or benign) or multiclass classification (identifying several categories of skin illnesses, such as cancer, benign lesions, and others) have shown positive results. Both of these classification methods are used to identify different types of skin disorders. In spite of the fact that the current detection rate is somewhere around 90%, there is still potential for improvement because relatively few studies have been carried out with the specific intention of detecting melanoma, which is a form of skin cancer. In addition, there has been no previous research that has examined the robustness of the models by using different datasets. This is because all of the studies have utilized the same dataset for both training and testing purposes. Furthermore, the Kaggle Skin Disease Dataset is utilized in this investigation for binary and multiclass melanoma classification. ResNet-50 and MobileNetV2 have been employed, and then HAM10000 is used to test the superior classifiers’ reliability. ResNet-50 and MobileNetV2 are trained with the Skin Disease Dataset and HAM10000 to detect melanoma better. Next, the classifiers are tested using ISIC2018, an independent test dataset. The results showed that ResNet-50 had the highest classification accuracy at 99.8%. This demonstrates that the binary classifier is superior to the multiclass classifier. Both binary models were only 50% correct when they were tested for robustness. On the other hand, ResNet-50 and MobileNetV2, both of which were trained on the pooled dataset, achieved a classification success rate of 98.8%. In contrast to the classification accuracy of 50% that was achieved through training with a single dataset, the robustness of ResNet-50 was tested, and the results showed that it achieved an accuracy of 85%.